Monte Carlo route optimization: score fixed plans, not per-sample routes
When asked to find the best agent route (e.g. a firefighter search order) by running a stochastic simulation many times, picking the best route per sample is clairvoyant: it uses knowledge of where the hazard is that the agent can't have. Instead, the decision variable must be a plan or policy fixed before the hazard is observed (route order plus reactive rules), and the agent's decisions should only read an Observation/belief struct, never the true sim state. Evaluate every candidate plan on the same K seeded samples (common random numbers) so differences are not noise. Generate candidates cheaply with Held-Karp DP over rooms using a hazard prior estimated from hazard-only runs, then re-score the top candidates with the full coupled simulation. If the simulation has no randomness, add seeded sampling of the inputs first.